【作者】 雷兰一菲；

【导师】 郎海涛；

【作者基本信息】 北京化工大学 ， 凝聚态物理， 2011， 硕士

【摘要】 目标识别和分类方法是当前光学、计算机视觉和人工智能等领域的研究热点,主要包括目标的表达和分类器的设计两个部分。近年来,基于局部图像特征的方法显著发展,本文就是以局部图像特征为基础,通过设计合适的目标表达模型,和适当的分类器,实现对特定目标的识别和分类。首先对局部图像特征进行了较为深入的研究,好的局部图像特征应具有重复率高、速度快和对图像变换的不变性,从上述三个方面对SIFT, SURF, Daisy等三种当今流行的典型局部图像特征进行了比较。提出了一种全新的图像点描述符(EPD),利用图像点周围特定窗口内采样像素点的颜色,梯度模值以及梯度方向构成的特征向量对图像点进行描述。将其与SIFT相比,在不变性方面与前者性能相近,但是其低维度在速度上更加具有优势。EPD描述符不仅适用于描述图像的极值点,而且适用于描述一般的图像点。这一特性使其可以用于稠密立体匹配,并显示了良好的匹配效果。对于局部图像特征的应用还可以延伸到伪造印章的识别领域,对于印文图像识别的配准难的问题,提出了全新的印章识别新体系,以随机生成的特征线一致性来判定印章的真伪,实验表明,这种方法适用于各种类型和内容的印章识别。在目标识别研究中,提取了局部图像特征后,采用BoW方法对这些特征向量聚类,形成特征分布直方图；然后采用SVM作为分类器,并且得到了相应的平均分类准确率；最终实现基于局部特征的目标识别和分类。更多还原

【Abstract】 Object recognition and classification is the hot issue in the fields of optical, computer vision and artificial intelligence etc.. It consists of object representation and classifier design. In recent years, the research of local image features developed fast. This paper presents object classification and recognition based on local image features.This dissertation studies local image features, a well designed local image features should possess three typical characteristics:repeatable, fast and invariant descriptor to image transformations. We surveyed and compared three popular and widely used local image features using novel evaluation metrics we presented.In this paper, we use neighborhood pixel characteristics, including HSV color space, Gaussian-weighted gradient magnitudes and orientations, sampled in specific window around interest point to enhance the description. Experimental results show that, the performance of EPD, in the distinctiveness and invariance aspects, is as good as SIFT, while the time cost of descriptor construction and matching is far less than it.Moreover, the EPD combines more image characteristics, which makes it be able to describe common image points, but not limited to the image extreme points. These advantages make the EPD finding new applications in the field of dense stereo matching, and it shows good match results.The application of local image features can also be extended to forged seal identification. The difficult problem of forged seal identification is image registration. We proposed the new system of seal identification, it can recognition by the consistency ratio of feature line randomly generated. The results show that this method is applicable to all types and content seals.The Research in object recognition, after extracting the local image feature, we can use BoW to cluster of these feature vectors in the form of histograms, then use support vector machine as classifier. The goal is object recognition and classification.更多还原